We have built a first version of our end-to-end AI/AGI robot brain that does planning and reasoning in many steps in a stable way. Our AI technology is not an LLM and does not use reinforcement learning. It learns like a child and has continual learning from feedback based on a structured world model, a prerequisite for true intelligence as described by the Turing prize Winner Yann LeCun and Nobel Prize Laureate Geoffrey Hinton.
We have benchmarked the performance of our A/AGII system with ACT, the Action Chunking Transformer (
https://arxiv.org/abs/2304.13705(odnośnik otworzy się w nowym oknie)) developed at Stanford in a cooperation with Google, used in the Aloha robot (Mobile ALOHA, Stanford Robotics Center) We measured the success rate (%) in a robot task where the robot arm and gripper interacted with a cube performing cube transfer. We demonstrated the lowest loss and fastest convergence, outperforming the baseline (ACT) with 80-100% success rate compared to 30-65%, and with 10-20 times lower loss depending on subtask and parameter setting.
We have successfully delivered an API that enables AI integration and control of industrial robots from the large robot OEMs like KUKA Robotics, ABB, Universal Robots. Our AI controls the robot all the way down to the joint angles, learning and executing trajectory, speed and acceleration and gripping force (not only picking point), which gives a higher level of flexibility in terms of handling objects of different sizes, weights and fragility.
